Anthropic turns Claude Mythos 5 loose on cyber defense
Three stories landed in the last cycle that point the same way: the most capable models are moving from lab benchmarks into real-world infrastructure, and the public is pushing back on the footprint that infrastructure leaves behind.
Anthropic is now powering its Claude Security scanner with Claude Mythos 5, its strongest model, to hunt vulnerabilities in enterprise codebases. The tool scans repositories for flaws, tags each finding with a CWE classification and severity rating, and suggests a patch — but a human still has to sign off on every fix. Mythos 5 is also being wired into partner security products that guard hospitals, utilities, and banks, so end users see results rather than the model itself. Anthropic has kept Mythos 5 tightly gated precisely because it excels at cyber tasks, and the bet here is to hand that edge to defenders without ever putting the raw capability in attackers' hands. It's a telling glimpse of where frontier labs think their models create the most value: not chat, but the unglamorous, high-stakes plumbing of software security.
DeepSeek released V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding to its fast V4-Flash text engine. On DeepSeek's own agent benchmarks, the vision variant nearly matches — and sometimes beats — Anthropic's Opus 4.8 on multimodal agent tasks, while keeping the base model's reasoning and world knowledge. It's built for visual agent workflows: it can describe images, pull text out of screenshots, and parse diagrams, and it speaks the same APIs as OpenAI and Anthropic so it drops into existing agent frameworks. Each image costs at most 384 tokens, a request can carry up to 600 of them, and pricing follows the cheaper V4-Flash tier. This is the clearest signal yet that DeepSeek is chasing the agentic, tool-using slice of the market rather than raw text bragging rights.
Public opposition to AI data centers in the US has nearly doubled in a year, jumping from 42 percent against to 75 percent, with 61 percent now "strongly opposed." A Heatmap News survey that asked the same question repeatedly over twelve months calls the shift "swift and massive," and the numbers track a Gallup poll from May that found 71 percent opposed. The top concerns are local power bills, water use, and land use — the unglamorous costs of the compute boom landing in people's backyards. For an industry racing to build out capacity, a three-out-of-four-not-in-my-town electorate is a permitting and political problem that won't be solved by better models.
What to watch: whether the data-center backlash pushes more frontier labs toward custom silicon and owned infrastructure, the exact path Anthropic and Waymo are already walking.
Which worries you more — the model defending the code, or the data center powering it?
Sources: The Decoder · Anthropic · The Decoder · DeepSeek · The Decoder · Heatmap News
